Greetings everyone. I am the speech analytics manager for our organization and I am currently working on fine-tuning our topic detection rules. We are seeing a high rate of false positives for our ‘Billing Dispute’ topic because the engine is triggering on any mention of the word ‘payment’ or ‘invoice’, even in a positive context. I want to add some proximity and exclusion rules to our topic definitions to ensure that we only flag interactions where there is actual conflict or dissatisfaction mentioned near those keywords. Does anyone have advice on the best way to structure these complex topic patterns?
Hello Ana20. I am a real-time dashboards developer and I have seen how these false positives can clutter our reporting. You should use the ‘Negative’ operators within your topic phrases. For example, you can use payment NOT (thank you) to exclude successful payment confirmations. Also, the ‘Proximity’ setting is very helpful; you can set it so that ‘payment’ must be within five words of ‘error’ or ‘refused’ to trigger the topic.
I use similar logic for our NLU intent training. In my experience, it is better to have several small, specific topics rather than one large, complex one. You could have one topic for ‘Payment Refused’ and another for ‘Invoice Discrepancy’. This makes it much easier to track the root cause of the billing issues in your analytics dashboards.
I maintain over fifty flows and I have seen these topics cause havoc with our automated routing if they are not tuned correctly. If you use topics to trigger a supervisor alert, a false positive will waste a lot of their time. Ana20, please make sure you use the ‘Topic Miner’ tool to test your new rules against your historical data before you make them active. It is the only way to see if your exclusions are actually working without waiting for a live call to fail.